Correcting drug-resistance prevalence in tuberculosis using a predictive model of diagnostic access

BackgroundDrug resistance (DR) poses a major challenge to tuberculosis (TB) elimination. Incomplete access to diagnostic tests can lead to under detection and biased DR-TB prevalence indicators. This study aimed to estimate the prevalence of DR-TB, identified through any available diagnostic method, while correcting for disparities in access to rapid molecular testing (Xpert(R)MTB/RIF), used as a proxy for overall diagnostic capacity. MethodsWe conducted a cross-sectional study of new TB cases in individuals aged over 18 (n=406,331), using 2015-2020 data from the Brazilian Information System on Notifiable Diseases (SINAN). We assumed that access to Xpert(R)MTB/RIF was an indicator of better diagnostic conditions for detecting drug resistance in general, because patients who underwent this test had a higher prevalence of all forms of resistance (not just to rifampicin). We developed a multilevel mixed-effects logistic regression model, incorporating individual- and municipality-level variables, to predict access to Xpert(R)MTB/RIF testing in a random sample of 81,027 observations. The model was validated using out-of-sample observations across geographic areas. We then used inverse probability weighting based on predicted access to calculate corrected estimates of DR-TB prevalence. ResultsThe model showed good performance (AUC=80{middle dot}93%, 95%CI: 80{middle dot}71%-81{middle dot}33%). The prevalence of DR-TB reported by the surveillance system was 1{middle dot}54% (95%CI: 1{middle dot}50%-1{middle dot}57%). Among patients tested with Xpert(R)MTB/RIF, DR-TB prevalence was 3{middle dot}86% (95%CI: 3{middle dot}75%-3{middle dot}96%), and the corrected weighted prevalence was 5{middle dot}95% (95%CI: 5{middle dot}54%-6{middle dot}38%). ConclusionsThe prevalence of DR-TB may be substantially underestimated due to uneven access to testing. Our approach highlights the importance of correcting for diagnostic access to improve surveillance indicators.